Memory and privacy

If I Delete Something, Is It Actually Gone?

> ANSWER[AEO]

Deleting from an AI memory system is harder than deleting a file, because these systems derive new artifacts — summaries, graph relationships, inferences — from the original fact, and removing the original does not automatically un-derive them. In Closer, deletion is scoped to the client: deleting a client erases that client's whole memory scope in the vector store and removes every note, transcript, action and knowledge-graph row Closer holds for them. There is no way to delete a single individual memory and leave the rest of the client intact.

Why isn't deleting an AI memory as simple as deleting a database entry?

Because modern memory systems do not store a fact once. They connect it to other facts, summarise it, and sometimes use it to generate further conclusions that get stored separately. Research on machine unlearning in agentic memory names the problem directly: "derived artifacts in memory (e.g., summaries, knowledge graph entities) may aggregate multiple sources, requiring dependency-aware deletion to avoid destroying shared artifacts."

In plain terms: if a fact was folded into a summary, or used to infer something else that was stored on its own, deleting the original does not undo what was built on top of it.

Is there a deeper problem than "we forgot to delete the copies"?

Yes, and it has a name in current research: information backflow. Researchers studying this in AI agents describe it precisely: "even after information is removed from the agent's memory, this parametric residue can then be used to regenerate the forgotten content during subsequent interactions, which is written back into memory and reverses the unlearning."

A system that successfully deletes something from its explicit memory store can, in some architectures, reconstruct that information later from patterns the underlying model already learned elsewhere — undoing the deletion with nobody intending it. The same work describes this as making "isolated unlearning strategies fundamentally insufficient" on their own.

Does that mean right-to-be-forgotten requests are impossible to honour?

Not impossible, but harder than most people assume, and better treated as an engineering commitment than a checkbox. What works, per the research, is designing for deletion from the start rather than retrofitting it: tracking dependencies between stored facts and anything derived from them, removing an original and everything meaningfully built from it together, and being explicit about the scope of what "delete" covers rather than implying a blanket guarantee the architecture cannot back.

What does deletion actually cover in Closer?

The specific answer, because a vague one is worse than none:

Does anything expire on its own?

No, not unless someone turns it on. Closer ships with the age-based retention sweep disabled. When an operator enables it, it deletes stored transcripts, cached transcripts and the memory audit log older than the configured cutoff — and it deliberately does not touch the vector store, because that store exposes no created-at filter to sweep on. That gap is documented in the code rather than papered over.

The practical consequence: if you want a client's data gone, delete the client. Do not assume time will do it.

What should you ask any product that claims deletion?

Key takeaways

  • > AI memory deletion is a live research problem: derived summaries and graph entities do not disappear when the original fact does.
  • > "Information backflow" describes a model regenerating deleted content from what it already learned, reversing an unlearning step.
  • > In Closer the unit of deletion is a client: their memory scope, notes, transcripts, actions and knowledge-graph rows all go together.
  • > Closer cannot delete one individual memory in isolation, and deletion is admin-only.
  • > Nothing expires automatically — the age-based retention sweep ships off, and it never purges the vector store even when enabled.

Frequently asked questions

If I delete a note or a client record, is the information actually gone from the AI's memory?

It depends on how deeply that information was used, and this is a genuinely unresolved technical problem across the industry, not something unique to any one product.

  • Current research identifies a specific failure mode called "information backflow".
  • Even after a fact is deleted from an explicit memory store, if it was ever summarized, combined with other facts, or used to generate other stored conclusions, traces of it can persist in those derived artifacts.
  • It can even be regenerated later from a language model's underlying weights.
  • How the write-and-retrieve path works: how AI memory actually works under the hood.
Why is this harder than just deleting a row from a database?

Because a knowledge graph and an AI memory system don't just store isolated facts — they connect facts to each other and often generate new, derived facts from the originals.

  • Summaries, inferences and relationship links are built on top of the original entry.
  • Deleting the original entry doesn't automatically un-derive everything that was built from it.
  • Researchers describe this as a "dependency-aware deletion" problem: you have to track and remove not just the fact itself, but everything downstream that depended on it.
  • More on the derived layer: what a client knowledge graph is.
So is 'delete this' just not a meaningful request for a system like this?

It's meaningful, but it's an engineering commitment, not a given — which is exactly why it's worth asking any product making this claim to be specific about what deletion actually guarantees.

  • A system built with deletion in mind from the start can make a real, honest guarantee: dependency tracking, no permanent model retraining on deleted content, clear scoping of what "delete" removes.
  • A system that treats deletion as an afterthought generally can't, however sincerely it intends to.
  • Worth asking of any vendor: what is the unit of deletion, does it reach the derived artifacts, and what happens if part of the delete fails.
What exactly does deleting a client remove in Closer?

Deletion is per client and admin-only — Closer has no way to delete one individual memory and leave the rest of the client in place. The blunt correction is deleting the whole client.

  • Deleting a client erases that client's entire memory scope in the vector store, then removes every row Closer owns for them in one transaction: notes and transcripts, pending actions, memory rows, and the knowledge graph entities and relationships built from those conversations.
  • The cached transcript keyed to a recording is purged too, unless another client's note still references the same recording.
  • The vector-store erase is best-effort by design: if that store is unreachable, the database delete still proceeds and the failure is only logged, so an operator has to check the logs and re-run the erase.
Does Closer delete old data automatically after a set period?

No, not by default. Closer ships with the age-based retention sweep switched off, so nothing expires on a timer unless an operator turns it on — and even when it is enabled it does not purge the vector store.

  • When enabled, the sweep removes stored transcripts, cached transcripts and the memory audit log older than the cutoff.
  • It leaves the vector store alone, because that store exposes no age filter.
  • Erasing memory vectors happens on client deletion, not on a schedule. If you want a client's data gone, delete the client — do not assume time will do it.

Sources